tao han
tao han
Before generating IIM from dot labels, we used the NWPU dataset to train a head scale prediction model, which estimates the box size for other crowd datasets. The pre-trained model...
Yes, we now have provided the code that can directly generate IIMs with points only. You can find the function `generate_masks_with_points()` in `dataset_prepare/prepare_NWPU.py`.
You can lower the learning rate of the threshold encoder in `config.py`, such as 1e-7. ``` if __C.OPT == 'Adam': __C.LR_BASE_NET = 1e-5 # learning rate __C.LR_BM_NET = 1e-7 #1e-6...
In our training, NaN would appear even if we lowered the threshold some times. At this time, we usually lower the threshold again to avoid this problem. We recommend using...
我们提供的是pytorch保存的模型,onnx格式的模型可自通过我们开源的模型参数自行转换
您好,我们刚才已经在SHHB数据集上重新验证了,对于训练过程中保存的模型是可以直接使用`test.py`进行预测的,请问您是不是clone的最新版本,以及是不是在后续改动过程中引起了新的问题? 训练过程中保存的模型:  在`test.py`中添加要测试的模型路径,以及修改数据集名称等相关参数:  在测试集上的预测结果:`./saved_exp_results/SHHB_HR_Net_test.txt'` 
你好,TE的主要目的不是去除背景,就像你说的,如果训练得好,背景区域基本已经没有噪声了,这种场景自然不需要TE。TE最主要的是想减少粘连的区域,例如,对于密集粘连区域,我们希望学出来的的阈值能够自适应的在这一块区域变大来实现更好的分割。
Thanks for your attention, I am sorry that I just see this issue. if lowing the initial lr doesn't work. I suggest that you can try to change the activation...
请确认val_gt_loc.txt和生成的预测文件的格式,如下: pred.txt 3110+人数 + 预测的人头坐标(x, y) 3111 ....... ............... val_gt_loc.txt 3110 + 人数 +人头的位置和size信息(x, y, sigma_s, sigma_l, level) 3111 ...... ................. 其中,3110,3111对应着NWPU数据集里图片的ID
see https://github.com/taohan10200/IIM/issues/28#issue-1164642630